際際滷

際際滷Share a Scribd company logo

HS Demo

Download as PPTX, PDF
Horace Sklar
Horace Sklar

The document discusses hyperspectral imaging and architectural trades for a hyperspectral imaging system. It describes how hyperspectral data is collected and formatted. It discusses potential architectural approaches including spatial scanning, spectral binning, and a SIMD processor array. It provides block diagrams of example hyperspectral imaging systems and payloads. It also discusses requirements and implementation considerations for mapping hyperspectral data to Landsat equivalent bands.

1 of 25
Conceptual Architecture Trades H.Sklar 2015
Hyperspectral Imaging Requirements Architectural Trades System Block Diagram Proposed Division of Labor
EO-1 Spatial scanning (Push Broom)
An airborne or spaceborne imaging sensor simultaneously samples multiple spectral wavebands over a large area in a ground-based scene. each pixel in the resulting image contains a sampled spectral measurement of reflectance, which can be interpreted to identify the material present in the scene.
Hyperspectral sensors collect information as a set of 'images Each image represents a narrow wavelength range of the electromagnetic spectrum, also known as a spectral band. These 'images' are combined to form a three-dimensional (x,y,了) hyperspectral data cube for processing and analysis, where x and y represent two spatial dimensions of the scene, and 了 represents the spectral dimension (comprising a range of wavelengths) The primary advantage to hyperspectral imaging is that, because an entire spectrum is acquired at each point, the operator needs no prior knowledge of the sample, and postprocessing allows all available information from the dataset to be mined.
Spatial scanning (Push Broom) NGC each two-dimensional (2-D) sensor output represents a full slit spectrum (x,了) obtain slit spectra by projecting a strip of the scene onto a slit and dispersing the slit image with a prism or a grating the spatial dimension is collected through platform movement or scanning. line-scan systems are particularly common in remote sensing Spectral scanning each 2-D sensor output represents a monochromatic ('single-colored'), spatial (x,y) map of the scene
Non-scanning a single 2-D sensor output contains all spatial (x,y) and spectral (了) data HSI devices for non-scanning yield the full datacube at once Spatiospectral scanning each 2-D sensor output represents a wavelength- coded ('rainbow-colored', 了 = 了(y)), spatial (x,y) map of the scene.
Hyperspectral imaging pushbroom spectrometers1,3,4 are currently used in several domains in order to identify the spectral signatures of a broad range of materials in the reflected solar energy spectrum. The camera images the scene line by line using the a so-called "pushbroom" scanning mode. The result can be seen as one 2d image for each spectral channel, or alternatively every pixel in the image contains one full spectrum. Spectrometers provide data under the form of hyperspectral cube. A hyperspectral cube with M across-track pixels, L alongtrack pixels, and P spectral bands is here considered. The plane formed by the across-track and the spectral dimensions is called frame; a frame has M spatial pixels (M columns) and P spectral pixels (P rows). Figure 1 shows how the sensor generates such a cube.
HS Demo
on-chip binning two or more spectral bands are summed up in a way that they form a unique row channel (Figure 2). This summation is done by the hardware during image acquisition. In general, the higher the number of binned rows (bands), the higher is the spectral SNR. Spectral binning will reduce the number of bands. frames are M x B matrixes, where B <= P. Current processing systems for the above sensor schemes incorporate a frame buffer that captures an image into memory. However, rapid detector-array advances in resolution, frame rate, and dynamic range will soon exceed throughput limits inherent in store-and process systems
The number of across-track spatial pixels is preserved. Whereas the bands (0,1,2) are binned to form band (0), bands (3,4) will form band 1 and so on. Landsat-7 Simulation uses Spectral Binning. In general, the higher the number of binned rows (bands), the higher is the spectral SNR.
Hyperspectral image-processing algorithms must be performed on many parallel PEs to maintain high throughputs. Rather than store the entire image frame the computation must be performed as the data arrive to minimize storage buffers.
Organization of a SIMD computer architecture. Program instructions are broadcast to every PE in the system through a single instruction stream, and each PE carries out the received instructions on its local data. P0, P1, Pn, PEs; MEM 0, MEM 1 MEM n, local memory.
Block diagram of the SIMD focal-plane system. Each PE in the SIMD processor array can address a 4 3 4 array of image sensors. An ALU with an addersubtractor and a barrel shifter. A multiplyaccumulate ~MACC! unit. Three-ported general-purpose register file and special register. Sixty-four words of local memory ~a maximum of 256 words!. Communication and serial IO units. A masking unit to control PE activity. This model permits the entire image as projected onto many PEs to be obtained in a single operation. Shift unit, barrel shifter; ADC
HS Demo
NASA's Earth Observing EO-1 with its hyperspectral instrument Hyperion implements Spatial scanning Hyperion Data is standard HDF Version 4.1 (v5) band-interleaved-by-line (BIL) files stored in 16-bit signed integer radiance values. Converted non-HDF format (off-line) so it is raw 16-bit signed, Little Endian Optionally unsigned 16-bit
Frame 256 pixels x 242 Bins (Frequency) (Push Broom) BIN 6 does not exist, 7 sets only Sensor 1 1-70 Bins; Multiplied by 0.025 Sensor 2 71 242 Bins; Multiplied by 0.0125 Freq Bin Coefficients Range: 0 - 12K counts 13.5 Bits 7 Bands of Proportional Data [242 x 7] Array of Numbers: Band Proportional Coefficients Dynamic Range = Log2 (12K-0) = 13.55 Bits 6K Frames of Data for Simulation Derived Rqmt (Miguel) (Need to Assess Architecture) Frame Rate ~ 60 Hz Implies pixel rate = 25 Mhz Clock Rate ~ 100 Mhz Functional (see Blk Dgm)
Floating Point IEEE754 vs Integer Number Conversion Operation Sizing
We are taking 242 Spectral Filters and mapping into the 7 Spectral Filters of LANDSAT-7 The LANDSAT Equivalent Pixel will then be 7 Rows INj,k x PCk,o = PDj,o LANDSAT Pixel Rows Resulting Image is a Frame of J x O or 256x7 Numbers Plus Average Band?
Design Schedule is TOP priority Need to show path to deliverable configuration Floating vs Integer MAC IEEE Floating point will let us get there faster Integer doable, but not now IEEE 754 standard specifies a binary32 as having: Sign bit: 1 bit Exponent width: 8 bits Significand precision: 24 bits (23 explicitly stored) Number Conversion Sequence Before or After Ping/Pong
IEEE574 Floating Pt Interger / Scaling
HS Demo
HS Demo
Miguel identified Alpha Data XRM-ZBT provides 2 banks of between 256K and 2048K x 36-bit ZBT pipelined memory 2 RS232 ports on the front panel
Partitioned with Resources in Mind Miguel HW and Interfaces with SATA & Display Horace System Architecture Issues, Trades Yogi VHDL Processing Engine

More Related Content

HS Demo

  • 2. Hyperspectral Imaging Requirements Architectural Trades System Block Diagram Proposed Division of Labor
  • 4. An airborne or spaceborne imaging sensor simultaneously samples multiple spectral wavebands over a large area in a ground-based scene. each pixel in the resulting image contains a sampled spectral measurement of reflectance, which can be interpreted to identify the material present in the scene.
  • 5. Hyperspectral sensors collect information as a set of 'images Each image represents a narrow wavelength range of the electromagnetic spectrum, also known as a spectral band. These 'images' are combined to form a three-dimensional (x,y,了) hyperspectral data cube for processing and analysis, where x and y represent two spatial dimensions of the scene, and 了 represents the spectral dimension (comprising a range of wavelengths) The primary advantage to hyperspectral imaging is that, because an entire spectrum is acquired at each point, the operator needs no prior knowledge of the sample, and postprocessing allows all available information from the dataset to be mined.
  • 6. Spatial scanning (Push Broom) NGC each two-dimensional (2-D) sensor output represents a full slit spectrum (x,了) obtain slit spectra by projecting a strip of the scene onto a slit and dispersing the slit image with a prism or a grating the spatial dimension is collected through platform movement or scanning. line-scan systems are particularly common in remote sensing Spectral scanning each 2-D sensor output represents a monochromatic ('single-colored'), spatial (x,y) map of the scene
  • 7. Non-scanning a single 2-D sensor output contains all spatial (x,y) and spectral (了) data HSI devices for non-scanning yield the full datacube at once Spatiospectral scanning each 2-D sensor output represents a wavelength- coded ('rainbow-colored', 了 = 了(y)), spatial (x,y) map of the scene.
  • 8. Hyperspectral imaging pushbroom spectrometers1,3,4 are currently used in several domains in order to identify the spectral signatures of a broad range of materials in the reflected solar energy spectrum. The camera images the scene line by line using the a so-called "pushbroom" scanning mode. The result can be seen as one 2d image for each spectral channel, or alternatively every pixel in the image contains one full spectrum. Spectrometers provide data under the form of hyperspectral cube. A hyperspectral cube with M across-track pixels, L alongtrack pixels, and P spectral bands is here considered. The plane formed by the across-track and the spectral dimensions is called frame; a frame has M spatial pixels (M columns) and P spectral pixels (P rows). Figure 1 shows how the sensor generates such a cube.
  • 10. on-chip binning two or more spectral bands are summed up in a way that they form a unique row channel (Figure 2). This summation is done by the hardware during image acquisition. In general, the higher the number of binned rows (bands), the higher is the spectral SNR. Spectral binning will reduce the number of bands. frames are M x B matrixes, where B <= P. Current processing systems for the above sensor schemes incorporate a frame buffer that captures an image into memory. However, rapid detector-array advances in resolution, frame rate, and dynamic range will soon exceed throughput limits inherent in store-and process systems
  • 11. The number of across-track spatial pixels is preserved. Whereas the bands (0,1,2) are binned to form band (0), bands (3,4) will form band 1 and so on. Landsat-7 Simulation uses Spectral Binning. In general, the higher the number of binned rows (bands), the higher is the spectral SNR.
  • 12. Hyperspectral image-processing algorithms must be performed on many parallel PEs to maintain high throughputs. Rather than store the entire image frame the computation must be performed as the data arrive to minimize storage buffers.
  • 13. Organization of a SIMD computer architecture. Program instructions are broadcast to every PE in the system through a single instruction stream, and each PE carries out the received instructions on its local data. P0, P1, Pn, PEs; MEM 0, MEM 1 MEM n, local memory.
  • 14. Block diagram of the SIMD focal-plane system. Each PE in the SIMD processor array can address a 4 3 4 array of image sensors. An ALU with an addersubtractor and a barrel shifter. A multiplyaccumulate ~MACC! unit. Three-ported general-purpose register file and special register. Sixty-four words of local memory ~a maximum of 256 words!. Communication and serial IO units. A masking unit to control PE activity. This model permits the entire image as projected onto many PEs to be obtained in a single operation. Shift unit, barrel shifter; ADC
  • 16. NASA's Earth Observing EO-1 with its hyperspectral instrument Hyperion implements Spatial scanning Hyperion Data is standard HDF Version 4.1 (v5) band-interleaved-by-line (BIL) files stored in 16-bit signed integer radiance values. Converted non-HDF format (off-line) so it is raw 16-bit signed, Little Endian Optionally unsigned 16-bit
  • 17. Frame 256 pixels x 242 Bins (Frequency) (Push Broom) BIN 6 does not exist, 7 sets only Sensor 1 1-70 Bins; Multiplied by 0.025 Sensor 2 71 242 Bins; Multiplied by 0.0125 Freq Bin Coefficients Range: 0 - 12K counts 13.5 Bits 7 Bands of Proportional Data [242 x 7] Array of Numbers: Band Proportional Coefficients Dynamic Range = Log2 (12K-0) = 13.55 Bits 6K Frames of Data for Simulation Derived Rqmt (Miguel) (Need to Assess Architecture) Frame Rate ~ 60 Hz Implies pixel rate = 25 Mhz Clock Rate ~ 100 Mhz Functional (see Blk Dgm)
  • 18. Floating Point IEEE754 vs Integer Number Conversion Operation Sizing
  • 19. We are taking 242 Spectral Filters and mapping into the 7 Spectral Filters of LANDSAT-7 The LANDSAT Equivalent Pixel will then be 7 Rows INj,k x PCk,o = PDj,o LANDSAT Pixel Rows Resulting Image is a Frame of J x O or 256x7 Numbers Plus Average Band?
  • 20. Design Schedule is TOP priority Need to show path to deliverable configuration Floating vs Integer MAC IEEE Floating point will let us get there faster Integer doable, but not now IEEE 754 standard specifies a binary32 as having: Sign bit: 1 bit Exponent width: 8 bits Significand precision: 24 bits (23 explicitly stored) Number Conversion Sequence Before or After Ping/Pong
  • 21. IEEE574 Floating Pt Interger / Scaling
  • 24. Miguel identified Alpha Data XRM-ZBT provides 2 banks of between 256K and 2048K x 36-bit ZBT pipelined memory 2 RS232 ports on the front panel
  • 25. Partitioned with Resources in Mind Miguel HW and Interfaces with SATA & Display Horace System Architecture Issues, Trades Yogi VHDL Processing Engine